Re: [Patch v4 12/16] x86, irq, ACPI: Implement interface to support ACPI based IOAPIC hot-addition

[Jiang Liu <jiang.liu@linux.intel.com>]

On 2014/9/9 20:20, Thomas Gleixner wrote:
> On Thu, 28 Aug 2014, Jiang Liu wrote:
>>  EXPORT_SYMBOL(acpi_register_ioapic);
>> diff --git a/arch/x86/kernel/apic/io_apic.c b/arch/x86/kernel/apic/io_apic.c
>> index 6e67af0c5f99..b286461cabf9 100644
>> --- a/arch/x86/kernel/apic/io_apic.c
>> +++ b/arch/x86/kernel/apic/io_apic.c
>> @@ -3851,7 +3851,13 @@ static int bad_ioapic_register(int idx)
>>  
>>  static int find_free_ioapic_entry(void)
>>  {
>> -	return nr_ioapics;
>> +	int idx;
>> +
>> +	for (idx = 0; idx < MAX_IO_APICS; idx++)
>> +		if (ioapics[idx].nr_registers == 0)
>> +			return idx;
>> +
>> +	return MAX_IO_APICS;
>>  }
>>  
>>  int mp_register_ioapic(int id, u32 address, u32 gsi_base,
>> @@ -3867,8 +3873,15 @@ int mp_register_ioapic(int id, u32 address, u32 gsi_base,
>>  	}
>>  	for_each_ioapic(ioapic)
>>  		if (ioapics[ioapic].mp_config.apicaddr == address) {
>> -			pr_warn("address 0x%x conflicts with IOAPIC%d\n",
>> -				address, ioapic);
>> +			/*
>> +			 * IOAPIC unit may also be visible in PCI scope.
>> +			 * When ioapic PCI driver's probe() is called,
>> +			 * the IOAPIC unit may have already been initialized
>> +			 * at boot time.
>> +			 */
>> +			if (!ioapic_initialized)
>> +				pr_warn("address 0x%x conflicts with IOAPIC%d\n",
>> +					address, ioapic);
> 
> Hmm. This smells fishy. Why do we allow multiple initializations of
> the same IOAPIC in the first place. Either it's done via ACPI or via
> PCI, but not both.
The ACPI subsystem will register and initialize all IOAPICs when walking
ACPI MADT table during boot, before initializing PCI subsystem.
Later when binding ioapic PCI driver to IOAPIC PCI device, it will try
to register the IOAPIC device again.

After this patchset is applied, we could remove the !ioapic_initialized
check. We check acpi_ioapic_register() before calling
acpi_register_ioapic(). So the check becomes redundant.
Or we could remove the temporary code from this patch.

> 
>>  			return -EEXIST;
>>  		}
>>  
>> @@ -3918,6 +3931,14 @@ int mp_register_ioapic(int id, u32 address, u32 gsi_base,
>>  	ioapics[idx].irqdomain = NULL;
>>  	ioapics[idx].irqdomain_cfg = *cfg;
>>  
>> +	if (ioapic_initialized) {
> 
> I have a hard time to understand this conditional. Why can't we do
> that unconditionally?
How about following comments?
/*
 * If mp_register_ioapic() is called during early boot stage when
 * walking ACPI/SFI/DT tables, it's too early to create irqdomain,
 * we are still using bootmem allocator. So delay it to setup_IO_APIC().
 */
Regards!
Gerry
> 
>> +		if (mp_irqdomain_create(idx)) {
>> +			clear_fixmap(FIX_IO_APIC_BASE_0 + idx);
>> +			return -ENOMEM;
>> +		}
>> +		alloc_ioapic_saved_registers(idx);
>> +	}
>> +
>>  	if (gsi_cfg->gsi_end >= gsi_top)
>>  		gsi_top = gsi_cfg->gsi_end + 1;
>>  	if (nr_ioapics <= idx)
> 
> Thanks,
> 
> 	tglx
>